app/routes/asr.py

from fastapi import APIRouter, UploadFile, File, HTTPException, Request
from faster_whisper import WhisperModel
import io
import asyncio

router = APIRouter()

# Global lock to protect RAM usage during concurrent requests
asr_lock = asyncio.Lock()

# Function to load the ASR model
def load_model_asr():
    # 'tiny' model -> minimal RAM usage, int8 -> even less memory
    model = WhisperModel("tiny", device="cpu", compute_type="int8")
    return model

@router.post("/asr")
async def asr(request: Request, audio: UploadFile = File(...)):
    """
    Transcribe audio to text using faster-whisper (locally, no HTTP calls)
    """
    async with asr_lock:
        try:
            # Get the local model from app.state
            model = request.app.state.model_asr

            # Read audio file into memory
            audio_bytes = await audio.read()
            buffer = io.BytesIO(audio_bytes)

            # Transcription
            segments, info = model.transcribe(buffer, beam_size=1, vad_filter=True)

            # Combine segments into full transcription
            transcription = " ".join([s.text for s in segments])

            return {"transcription": transcription}

        except Exception as e:
            # Raise HTTP exception if anything goes wrong
            raise HTTPException(status_code=500, detail=str(e))